LARC Lifts BAGEL to 0.88 GenEval via Continuous Latent Reasoning

LARC, a reasoning framework for Unified Multimodal Models described in a paper on Hugging Face, posts 0.88 overall on GenEval when applied to BAGEL. That is a seven-percentage-point gain over the base model, with the biggest jumps on attribute binding (+15 points) and spatial position (+13 points).

The authors' framing is that reasoning has failed to help image generation because the two live in different spaces. "We attribute this failure to a fundamental modality mismatch: reasoning is performed in a discrete token space, whereas image generation operates in a continuous space," the paper writes. LARC's answer is to reason over continuous hidden states rather than text tokens, interleaving language-mode and latent-mode steps.

Training runs in two stages. A curriculum SFT phase "gradually converts token-level reasoning into latent reasoning," and an information-gain RL stage then "self-evolves latent reasoning without ground truth supervision," using GRPO with an information-gain reward and a collapse-prevention term for diversity. The paper reports LARC also lifts prompt following on DPG-Bench to 90.04 from 85.07, and visual reasoning on MIRA to 11.7 from 7.5.

The headline claim is scaling behavior. The abstract says LARC "demonstrates clear test-time scaling behavior, outperforming token-based generation under both parallel and sequential test-time scaling strategies," with doubled latent compute yielding monotonic gains and matched-budget comparisons outperforming best-of-N sampling. Those numbers come from a single paper on a single base model, and the result lands amid a busy stretch of multimodal releases on our multimodal tracker. No third-party reproduction has appeared yet.